autotranslate translates through nr-llm

Context

A site that wants its page trees translated in the background, only where a translation is missing, again when the source changes, switched per table and field, with news and file references along, needs a translation workflow. The third-party extension thieleundklose/autotranslate is that workflow: a queue with a backend module and a console command, a save hook, and per-table settings in the site configuration. nr-llm has the parts it lacks: a translation service that reaches DeepL or a language model with the site's glossary, cost tracking and budgets. A separate extension that rebuilds the workflow on nr-llm was considered and not chosen.

autotranslate builds a \DeepL\Translator in one static helper and offers no way to replace it. An installation that wanted nr-llm instead patched that helper to resolve a class named in autotranslate's configuration from the container. Its only event, AfterTranslateEvent, fires after the translation, before autotranslate stores the translated columns with plain SQL.

The second gap is the notice. nr-llm marks the records its own runs write (ADR-239) and labels them in b13/ai-label, but autotranslate writes past both.

Decision

autotranslate keeps the workflow; nr-llm supplies the translator. nr-llm ships AutotranslateTranslator, a \DeepL\Translator subclass, as a public service. autotranslate uses it when its setting translatorClass names the class. That setting is offered upstream as a configuration option (autotranslate pull request 175), the mechanism of the patch made official, so that nr-llm depends on no unreleased autotranslate code: it works with the patch and with a release that carries the option alike.

The translator overrides what autotranslate calls. translateText() sends the texts of one call to TranslationServiceInterface::translateBatchWithTranslator() in one batch, a blank text excepted. DeepL's tag_handling becomes nr-llm's tag handling, so rich text keeps its markup. A target code keeps its region in nr-llm's form (EN-GB becomes en-GB), which nr-llm's DeepL translator needs; a source code carries none. The nr-llm translator is the setting autotranslate.translator; empty leaves the choice to the service, which uses the language model. getUsage() reports no quota, because nr-llm keeps its own budgets. listGlossaries() finds none, because a DeepL glossary id means nothing to the language model. The language lists are DeepL's codes, which autotranslate stores in the site configuration. The parent class is built with a placeholder key and a server under the reserved .invalid domain, so a method the class does not override fails without sending a text to DeepL. A translator that answers fewer texts than it was given fails the call: a missing answer would move every later text into the wrong field. So does a blank answer, an answer of markup without visible text for a text that has some, and one cut off at the output limit (truncated, ADR-209): autotranslate stores what it gets, and a half text in an unattended batch would stand as the translation.

On the command line the key is read as the worker. autotranslate's batch command is a console worker. It reads the provider key as worker.technicalBeUserUid, the identity ADR-234 gives nr-llm's own queue worker; a web request keeps its user. The integration reads the setting itself and does not use the agent module's scope class, because a translation component must not depend on the agent module (ADR-090).

A machine-translated record is marked in the write that stores it. Two listeners on AfterTranslateEvent add columns to the translated ones: the marker tx_nrllm_ai_written on pages and content elements, and, when AI Label is loaded and enables the table, its tx_ailabel_metadata. The label follows the rule of nr-llm's own translations: AI-modified, an AI-generated origin kept, the human review cleared. Both apply only while translatorClass names nr-llm's translator, and only when a text column (input, text, flex) holds text the translation changed from the source; the fields autotranslate copies unchanged do not count. A failed translation dispatches no event and stores neither.

Registered only for an active autotranslate. A compiler pass registers the integration when autotranslate's batch command is a service, and the AI Label listener when AI Label's API is one, as for AiLabelCompilerPass.

Consequences

• Background runs, retranslation on change, per-field settings, news and file references stay autotranslate's, with its limits: a batch translates every configured field again, and a change of the source overwrites a hand correction in the changed field.

• Until autotranslate releases the translatorClass option, an installation carries it as a patch. nr-llm does not change when the release arrives.

• The label is written with plain SQL in autotranslate's write, not through AI Label's API. A write through the API would be a second DataHandler run with its own history entry and save hooks, made before the text it labels is stored. It therefore leaves no history entry, as autotranslate's text write leaves none.

• In the backend a translation routed through nr-llm counts for the editor who started it. On the command line the worker identity only reads the key; the usage is recorded for the backend user of the process, _cli_, as for any other nr-llm call there.

• A failed translation leaves the record autotranslate's localization created, holding the copied source text, as a DeepL error does.

• nr-llm's site glossary (ADR-208) does not apply: translateText() carries no page, so the translator cannot tell the site.

• A translation that autotranslate makes with DeepL itself is not marked.

Alternatives considered

A new extension on nr-llm. It would own the workflow end to end, at the price of rebuilding a queue, a module, a context menu, a save hook and per-table settings that autotranslate maintains already.

A PSR-14 event upstream that hands over a translator. It would avoid a class name in the configuration. nr-llm would then depend on an event class that no autotranslate release contains yet, and could not be tested or shipped before that release.

Marking through nr-llm's provenance event. AfterAiRecordWrittenEvent names an agent run and is dispatched after the write. An autotranslate translation has no run, and AfterTranslateEvent comes before the write.

Revisit when

• autotranslate passes the page or site to its translator, so the site glossary can apply.

• autotranslate offers an event or an interface for the translator instead of a class name.